TAM-RL: Task-Aware Modulation in Carbon Flux Upscaling

Updated 5 July 2026

The paper introduces TAM-RL, a novel framework that couples task-conditioned modulation with physical constraints to upscale terrestrial carbon fluxes and reduce RMSE by up to 9.6%.
It leverages a shared BiLSTM encoder and an LSTM decoder with FiLM-based modulation to convert sparse eddy-covariance observations into continuous flux fields.
The system enforces carbon balance and non-negativity constraints, resulting in higher explained variance and robust zero-shot transfer across diverse biomes.

Task-Aware Modulation with Representation Learning (TAM-RL) is a framework that combines a shared representation learner with task-conditioned modulation, most explicitly instantiated for terrestrial carbon-flux upscaling as a system that couples spatio-temporal representation learning with a knowledge-guided encoder-decoder architecture and a loss derived from the carbon balance equation, $\mathrm{NEE} = \mathrm{RECO} - \mathrm{GPP}$ (Rozanov et al., 10 Mar 2026). In that formulation, TAM-RL is designed to infer spatially continuous, temporally resolved flux fields from sparse eddy-covariance observations while improving robustness under domain shift. Across more than 150 flux tower sites representing diverse biomes and climate regimes, it was reported to reduce RMSE by 8–9.6% and to raise explained variance from 19.4% to 43.8%, depending on the target flux (Rozanov et al., 10 Mar 2026).

1. Problem setting and motivation

TAM-RL was introduced to address the upscaling of terrestrial carbon fluxes from sparse eddy covariance tower measurements and dense gridded predictors. The principal targets are Gross Primary Production (GPP), Ecosystem Respiration (RECO or $R_{\mathrm{eco}}$ ), and Net Ecosystem Exchange (NEE). In the ONEFlux sign convention used in the work, NEE is negative for uptake and positive for release, while GPP and RECO are typically non-negative (Rozanov et al., 10 Mar 2026).

The methodological motivation is the combination of sparse and regionally biased tower coverage with strong domain shift. The cited work emphasizes that flux-tower observations are concentrated in North America and Europe, while many biomes and climate regimes remain underrepresented. Under such conditions, models trained on limited domains can exhibit systematic regional biases and high predictive uncertainty. The paper further argues that satellite indices under different meteorological regimes can correspond to different physiological states, so purely statistical models may generalize poorly beyond observed domains (Rozanov et al., 10 Mar 2026).

Within this framing, TAM-RL is presented as a response to three limitations of common upscaling practice. Classical spatial interpolation ignores ecosystem processes and temporal variability. Tree-based and purely data-driven models such as XGBoost often depend on extensive manual features and can overfit to observed domains. Existing data-driven products may therefore fail to generalize beyond the training support. TAM-RL addresses these limitations by learning spatio-temporal representations intended to capture process-level dependencies, by adapting shared representations to site-specific context through task-aware modulation, and by enforcing physically grounded constraints through the carbon balance relation (Rozanov et al., 10 Mar 2026).

2. Architectural formulation

In the carbon-flux formulation, TAM-RL combines a modulation network with a shared temporal decoder. The modulation network comprises a task encoder $E$ and a generator $G$ , while the forward model $F$ is an LSTM decoder that predicts target fluxes from driver sequences (Rozanov et al., 10 Mar 2026).

Component	Implementation	Role
Task encoder $E$	BiLSTM	Encodes support sequences into site embedding $z_i$
Generator $G$	MLP	Produces FiLM parameters from $z_i$
Forward model $F$	LSTM decoder	Predicts flux sequences from driver sequences

The task encoder ingests support sequences for a site $R_{\mathrm{eco}}$ 0 and outputs a task embedding $R_{\mathrm{eco}}$ 1 that captures temporal and ecosystem characteristics. The generator maps this embedding to feature-wise linear modulation parameters, and these parameters are applied both to the decoder input and to the final hidden state. The core modulation equations are

$R_{\mathrm{eco}}$ 2

where $R_{\mathrm{eco}}$ 3 are generated from $R_{\mathrm{eco}}$ 4 and $R_{\mathrm{eco}}$ 5 denotes element-wise multiplication (Rozanov et al., 10 Mar 2026).

The representation-learning substrate is built from daily sequences of remote sensing, meteorology, and static/contextual features. Inputs include MODIS surface reflectance products MOD09GA, land-cover MCD12Q1, ERA5-Land meteorological drivers, and contextual variables such as IGBP class and Köppen–Geiger climate type. Temporal structure is handled with 45-day sequences and a 15-day stride, harmonized at daily resolution to capture intra-seasonal dynamics (Rozanov et al., 10 Mar 2026).

Training proceeds in two stages. First, the decoder is pretrained on pooled data without modulation so that it learns a robust shared representation across sites. Second, the full system is trained jointly with support/query splits per task, with gradients backpropagated through the BiLSTM encoder, the MLP generator, and the LSTM decoder. This enables site adaptation without test-time weight updates, which the paper characterizes as zero-shot transfer. At inference on a new site, historical data are used as support to compute $R_{\mathrm{eco}}$ 6 and the corresponding FiLM parameters, but the network weights themselves are not fine-tuned (Rozanov et al., 10 Mar 2026).

3. Physically grounded objective and multi-flux coupling

A defining feature of TAM-RL in the carbon-flux setting is the use of physically grounded constraints in the loss. The governing relation is

$R_{\mathrm{eco}}$ 7

with the standard EC/ONEFlux convention that $R_{\mathrm{eco}}$ 8 indicates net uptake and $R_{\mathrm{eco}}$ 9 indicates net release (Rozanov et al., 10 Mar 2026).

The training objective combines a weighted supervised error with a flux-consistency penalty. The weighting scheme uses a continuous quality-control weight $E$ 0 derived from $E$ 1, together with inverse-frequency weights $E$ 2 and $E$ 3 to counter class imbalance across land-cover and climate classes. The physics term is

$E$ 4

and the scalar coefficient on this term is reported as $E$ 5 (Rozanov et al., 10 Mar 2026).

This objective induces multi-task coupling across the three flux outputs. Even when only GPP and NEE are directly supervised, RECO is produced as a latent output guided by the carbon balance. The framework also enforces non-negativity of GPP and RECO at inference by clipping: $E$ 6 The paper interprets this combination of weighted supervision, carbon-balance regularization, and constrained outputs as a mechanism for improving consistency and transferability across heterogeneous sites (Rozanov et al., 10 Mar 2026).

4. Data sources, training protocol, and evaluation design

The empirical study uses 579 sites from FLUXNET, AmeriFlux, ICOS, and JapanFlux spanning 2000–2023 and covering all Köppen–Geiger climate classes. Target variables are ONEFlux daily $E$ 7 and $E$ 8, while RECO is generated by the model and guided through the physics constraint. Remote-sensing features are extracted from MODIS MOD09GA and MCD12Q1 over a $E$ 9 window around each tower using Google Earth Engine, and meteorological predictors are taken from ERA5-Land at daily resolution (Rozanov et al., 10 Mar 2026).

The site split distinguishes training sites used in FLUXCOM-X-BASE from held-out test sites that are not in the FLUXCOM-X-BASE training list. The held-out out-of-domain evaluation includes 164 test sites. Reported evaluations are performed on highest-quality observations with $G$ 0, and are also reported with $G$ 1 when available. All inputs are aggregated to daily resolution and arranged into 45-day windows with 15-day stride (Rozanov et al., 10 Mar 2026).

The baseline set is chosen to isolate the contribution of temporal modeling and of task-aware modulation.

Baseline	Characterization	Role
FLUXCOM-X-BASE	XGBoost upscaling product aggregated to daily	external state-of-the-art reference
XGBoost	Flattened, non-temporal features	purely data-driven baseline
CT-LSTM	Standard LSTM with one-hot static features	temporal baseline
TAMLSTM	Decoder-only component of TAM-RL	modulation ablation

The paper states that the decoder is pretrained before joint training, that the support encoder is a BiLSTM and the query model an LSTM decoder, and that each neural-network configuration is averaged over 10 independent runs with different random seeds. It also explicitly notes that optimizer, learning-rate schedule, batch size, and number of epochs are not reported. The resulting ensemble of seeds provides a rudimentary uncertainty estimate, but no explicit probabilistic output is produced (Rozanov et al., 10 Mar 2026).

5. Quantitative performance, ablations, and failure modes

On held-out sites evaluated with $G$ 2, TAM-RL improves over FLUXCOM-X-BASE for both GPP and NEE. The reported average site-level values are as follows (Rozanov et al., 10 Mar 2026).

Target	TAM-RL	FLUXCOM-X-BASE
GPP	RMSE 1.97, $G$ 3 0.43	RMSE 2.18, $G$ 4 0.36
NEE	RMSE 1.62, $G$ 5 0.23	RMSE 1.76, $G$ 6 0.16

For GPP, the paper reports a 9.6% RMSE reduction and a 19.4% increase in explained variance, corresponding to an absolute $G$ 7 increase of 0.07. For NEE, it reports an 8.0% RMSE reduction and a 43.8% increase in explained variance, again with an absolute $G$ 8 increase of 0.07. Per-site scatter plots are described as showing lower RMSE and higher $G$ 9 than FLUXCOM-X-BASE across most sites and across diverse climate types (Rozanov et al., 10 Mar 2026).

The biome- and climate-stratified analysis indicates that CT-LSTM, TAMLSTM, and TAM-RL consistently outperform XGBoost and FLUXCOM-X-BASE across the five dominant Köppen–Geiger climate types. However, degradation persists for certain IGBP classes such as water bodies (WAT), and moderate drops are reported for MF, DBF, and ENF. A heatmap of relative RMSE similarly indicates comparative gains across IGBP types, with sample counts shown per type (Rozanov et al., 10 Mar 2026).

Ablations attribute a substantial fraction of the improvement to task-aware modulation itself. TAMLSTM, which removes the modulation network while retaining the decoder, is reported to be comparable to CT-LSTM; adding the modulation network yields the lowest RMSE and highest $F$ 0 for both GPP and NEE. The physics-informed loss is not isolated in a separate ablation, so its contribution is not quantified independently. The paper nevertheless states that the observed improvements, together with enforced carbon balance and non-negativity, suggest gains in physical consistency and transferability. Reported limitations include lower performance in underrepresented ecosystems, likely sensitivity to unmodeled disturbances such as drought and fire, possible confusion from inconsistent NEE sign conventions across datasets, and limited uncertainty quantification beyond seed ensembling (Rozanov et al., 10 Mar 2026).

The literature provided here suggests that TAM-RL has both a narrow and a broader usage. In a narrower sense, it denotes the carbon-flux upscaling framework described above. In a broader sense, it refers to architectures in which a shared representation is adapted by task-conditioned modulation rather than by full task-specific retraining (Rozanov et al., 10 Mar 2026).

An earlier paper titled "Task Aware Modulation using Representation Learning: An Approach for Few Shot Learning in Environmental Systems" describes TAM-RL as a multimodal meta-learning framework for few-shot learning in heterogeneous systems, with a modulation network and a base network trained through an amortized process. In the abstract, that work reports an 18.9% RMSE improvement over MMAML on FLUXNET with one month of few-shot data, an 8.21% improvement for streamflow prediction with one year of data, and at least 3x faster training times than gradient-based meta-learning approaches (Renganathan et al., 2023).

Several adjacent papers instantiate the same general principle with different mechanisms and domains. "Active Representation Learning for General Task Space with Applications in Robotics" formalizes task-aware and task-agnostic representation learning and explicitly discusses TAM-RL in terms of shared trunks, task embeddings, and modulators such as FiLM, gating, and hypernetworks; in its bilinear theory, the sample complexity for target-aware active representation learning scales as $F$ 1 (Chen et al., 2023). "Task-Aware Information Routing from Common Representation Space in Lifelong Learning" realizes task-aware modulation as task-specific gating masks generated by undercomplete autoencoders, applied to a common representation space in continual learning (Bhat et al., 2023). "Task adaption by biologically inspired stochastic comodulation" uses stochastic gain modulation targeted to task-relevant channels and neurons, emphasizing low-dimensional multiplicative noise as a task-dependent control signal in multi-task learning (Boeshertz et al., 2023).

Related formulations also appear outside supervised environmental prediction. "Task-Induced Representation Learning" studies task-induced representations for reinforcement and imitation learning in visually complex environments and reports that task-induced representations can double learning efficiency compared to unsupervised alternatives (Yamada et al., 2022). "HyperTASR: Hypernetwork-Driven Task-Aware Scene Representations for Robust Manipulation" uses a hypernetwork to generate task- and phase-conditioned transformation parameters for scene representations in robotic manipulation (Sun et al., 26 Aug 2025). "Task-Aware Adaptive Modulation: A Replay-Free and Resource-Efficient Approach For Continual Graph Learning" freezes a GNN backbone and inserts per-task Neural Synapse Modulators selected through a prototype-guided strategy, reporting 0.0% average forgetting across six GCIL benchmark datasets (Liu et al., 31 Aug 2025).

Taken together, these works indicate that TAM-RL is best understood not as a single architectural template but as a design family: a shared representation learner is preserved, while task context modulates internal computation through FiLM parameters, gates, stochastic gains, hypernetwork-generated weights, or related mechanisms. In the terrestrial carbon-flux setting, this family is specialized by domain constraints from carbon balance, site-conditioned support/query adaptation, and zero-shot transfer across held-out ecosystems (Rozanov et al., 10 Mar 2026).